It helps to think of fights against monsters as a turn-based encounter. As long as you can dodge or the monster misses its attack, you should be able to land a hit. If you get hit or are too far away when the monster attacks, you probably won’t be able to land any meaningful offense or heal without getting punished for it.
You can run language models on consumer cards right now. The only thing is depending on the size of the model and the amount of VRAM on your card, you might not be able to do much else.
You keep moving the goal posts and putting words in my mouth. I never said you can do new things out of nothing. Nothing I mentioned is approaching, equaling, or exceeding the effort of training a model.
You haven’t answered a single one of my questions, and you are not arguing in good faith. We’re done here. I can’t say it’s been a pleasure.
But at what point does that guidance just become the dataset you removed from the training data?
The whole point is that it didn’t know the concepts beforehand, and no it doesn’t become the dataset. Observations made of the training data are added to the model’s weights after training, the dataset is never relevant again as the model’s weights are locked in.
To get it to run Doom, they used Doom.
To realize a new genre, you’ll “just” have to make that game the old fashion way, first.
Or you could train a more general model. These things happen in steps, research is a process.
There are more forms of guidance than just raw words. Just off the top of my head, there’s inpainting, outpainting, controlnets, prompt editing, and embeddings. The researchers who pulled this off definitely didn’t do it with text prompts.
I mean, you’ve never seen a purple elephant with a tennis racket. None of that exists in the data set since elephants are neither purple nor tennis players. Exposure to all the individual elements allows for generation of concepts outside the existing data, even though they don’t exit in reality or in the data set.